First, the predictions themselves, in black and white, precise to the year:

Ten years from now, 2036: most companies take an AI-native form — the few people left in the organization set goals and verify results, while the execution is done by a swarm of AI Agents; robots have widely entered ordinary people’s work and lives, and offline industries like supermarkets, retail, and restaurants have been rewritten at scale.

Twenty years from now, 2046: robots are commonplace in homes; in intelligence, learning, manipulation, and movement, they fully surpass humans.

For a prediction to deserve “precise to the year,” it has to lay out the reasoning. Below is the complete five-step reasoning chain — every step backed by measured data, every step backed by evidence that has already happened. Work through the chain and the years grow out of it on their own; I didn’t invent them.

一、Step One: What Is AI’s Measured Performance Growth Rate?

No feelings, just measurement. METR, a US organization that specializes in benchmarking AI’s ability to work autonomously, has tracked the same metric for six years: how much human-expert work time corresponds to a task an AI can reliably complete.

In 2023, GPT-4: 4 minutes.
In 2025: 5.9 hours.
In April 2026, the latest frontier models: 17.4 hours.

Three years, more than 200x, doubling every 7 months on average, without pausing a single year in six.

You might suspect this is one company’s luck. Here’s an independent line of evidence: real-computer-use benchmarks (having AI use a desktop like a human to complete work tasks). In early 2025, the best AI scored one-sixth of human level; in 2026, it surpassed humans. In one year, from far behind to overtaking.

Beneath this curve is physical infrastructure: training compute growing 5x per year, global installed compute growing 3.4x per year, and the price of reaching equivalent capability falling 47% per quarter — 18 months ago one PhD-level science problem cost about $0.04, today it costs about $0.0006, a 725-fold drop. For comparison: electricity took seventy years to get this cheap.

Remember these three numbers: doubling every 7 months, halving in price every quarter, the base growing 3–5x per year. This is the foundation of the reasoning chain.

二、Step Two: The Growth Won’t Stop at 7 Months — RSI Will Make “Getting Smarter” Itself Get Faster

Some will say: doubling has to end sometime. That is precisely the key fork, and the reason I dare extrapolate twenty years: AI is being used to accelerate AI’s own development, and this has already been measured.

At the most advanced AI labs, per-engineer code output is 8x what it was a few years ago — and a substantial share of that code goes into building the next generation of models. Independent estimates of “how much AI speeds up AI R&D”: currently about 1.5x — AI can finish a year and a half of R&D work in one year.

1.5x sounds unsexy. Run the compounding anyway:

1.5x per year is 57x over ten years. If it rises to 2x, ten years is 1024x.

That is to say, even if the “1.5x speedup” figure never rises again, the effective intelligence invested in AI R&D in 2036 will be dozens of times today’s, and thousands of times by 2046. And this loop is self-reinforcing: stronger models → faster R&D → even stronger models. A positive slope isn’t scary; what’s scary is that the slope of the slope is also positive.

This is the popular version of RSI (recursive self-improvement). It hasn’t yet reached “AI fully autonomously building AI” — today humans set direction and AI does the work — but compounding doesn’t care about labels, only about the multiplier.

三、Step Three: Why RSI Can Industrialize “Discovery” Itself — The Search Paradigm

The speedup from Step Two acts on an entirely new way of solving problems. Today’s AI conquers hard problems not through some genius’s flash of insight, but:

Set a hard goal + build an automatic verification environment + unleash a swarm of Agents to brute-force search around the clock.

The power of this paradigm is proportional to how cheap verification is. Look at what has already happened:

Mathematics, verified by formal proofs (verification is free) — in October 2026, OpenAI published 722 papers at once, solving a batch of open problems that had hung unsolved for thirty years. A PhD grinds on one problem for a lifetime; a swarm grinds through a whole research direction in one night.

Biology, verified by physics models plus wet experiments — Claude discovered an entirely new class of enzyme systems this September.

Proteins, verified by structure prediction — AlphaFold in three years cleared structural biologists’ fifty-year backlog, and won a Nobel Prize.

After mathematics, code, and proteins, the next fields where verification cost is being hammered down are: materials, chemistry, chip design, manufacturing processes. Self-driving labs (robots run experiments, AI proposes hypotheses, iterating in a loop) are compressing “synthesize a new substance” from months to days. Battery energy density, motor materials, tactile sensing — the hard bottlenecks of the robot body are exactly the kinds of problems the search paradigm can attack.

This step is the heart of the entire prediction: for the first time, the speed of technological progress depends not on the number of scientists but on the throughput of compute and verifiers. And compute is growing 3–5x per year.

四、Step Four: How This Force Transmits to Robots — The Bottleneck Is the “Brain,” and the “Brain” Is Software

The robotics world has a saying, and two independent sources put it identically. UBTech’s vice president: “Without a strong brain, the body is just expensive hardware.“ Third-party analysis: “Hardware currently leads software; the brain is still in its infancy.”

Which means: robots’ shortfall today is not the body, it’s the mind — perception, planning, manipulation intelligence. And the mind is software. Software happens to be the home turf of the curves in steps one, two, and three.

The body side is moving too, with considerable acceleration: Chinese robot whole-unit costs, thanks to supply-chain localization, are falling 16% per year (localization of harmonic reducers already exceeds 75%) — and this is before RSI has been massively thrown at materials science. The two legs walk together:

• The brain approaches “reliably working in unfamiliar environments” at software speed (doubling every 7 months);
• The body approaches “cheaper to buy a machine than hire a person” at an accelerating hardware rate (costs down 16% a year → faster).

The industry’s judgment on the “embodied-AI ChatGPT moment” (completing 80% of tasks in unfamiliar environments): two to three years at the fastest, five to ten at the latest. Take the middle: around 2030. From “able to work” to “spread across every industry,” add one equipment-replacement cycle — that’s where 2036 comes from. It wasn’t slapped down, it was derived.

五、Step Five: Why China Will Run Fastest — Demographics Clear the Obstacles for Automation

The same technology landing on different demographic structures moves at completely different speeds.

Europe and America’s first reaction to automation is “they’re taking our jobs” — union negotiations, legislative restrictions, strikes (Korea already saw the first strike against humanoid robots). China faces a different arithmetic: the working-age population is shrinking by several million every year, and delivery, elder care, and factory front lines already can’t hire.

So in China, robots are not replacement but gap-filling — and that puts policy direction, capital flows, and social tolerance all on automation’s side. Evidence already in hand: L4 autonomous delivery vehicles added 30,000 units in a year, growing 350%, covering more than 300 cities; robotaxis have surpassed 23 million cumulative orders, and the Wuhan fleet is already profitable; UBTech’s ten-thousand-unit-scale humanoid robot factory began production this September, and in October signed to deploy on an automaker’s production line.

The five-step chain closes here, strung into one sentence:

Compute multiplies every year (the foundation) → RSI compounds R&D speed (the accelerator) → the search paradigm industrializes discovery (the method) → robots’ weakness happens to be software (the transmission) → China’s demographic gap clears deployment obstacles (the landing).

Reading the years off this chain: cognitive tasks cross the line around 2030; from structured settings (warehouses, long-haul routes, commercial kitchens) spreading to semi-open settings (supermarkets, restaurants, storefronts), one equipment cycle, 2036; unstructured environments plus “reliability must be verified by the calendar” (a machine meant to accompany an elderly person for twenty years — its safety certification cannot be brute-force searched; this is the one thing in the physical world algorithms cannot compress), add ten more years — 2046.

There, the reasoning is done. Now let me show you concrete life in those two years.

六、2036: Companies and Street Level Get Rewritten First

Companies first. Most companies in 2036 look like this: one business unit, three to five humans — setting goals, accepting deliverables, signing and taking responsibility — with an entire swarm of Agents hanging beneath them. You don’t manage colleagues; you manage a fleet of AIs. The concept of the “department” disappears, because a department’s function is to coordinate humans, and AI doesn’t need coordinating — it needs goals.

This transfusion has already begun, and there’s a counterintuitive detail: AI’s first wave doesn’t make headcount shrink, it makes output grow and execution get cheaper. What actually disappears is something else — entry-level jobs. Stanford’s data is already out: among US software developers aged 22–25, employment has fallen nearly 20% from peak, while senior staff over 35 rose instead of falling during the same period. Not layoffs — simply no longer hiring — the work newcomers did, AI now does. The bottom three rungs of the ladder have already been removed. By 2036 this will grow into a full generational dividing line: companies either want “people who can directly lead AI teams” (you have no experience) or “manual labor cheap enough not to be worth automating.” The middle layer is canceled wholesale.

Now the street. You walk downstairs: an autonomous delivery vehicle is parked at the supermarket’s receiving dock; shelf restocking is done by machines, the kitchen is a cluster of cooking machines and sorting arms, and the one or two “humanoid waiters” out front are mostly a publicity prop — the machines actually doing the work are the clusters in the kitchen and the warehouse, not the one bowing to your face. This National Day holiday, the news of humanoid robots selling ice cream at scenic spots in seven cities, 500 cups per unit per day, was a preview of exactly this picture — except ten years from now, they move from tourist sites into street-level shops.

七、2036: Home and School Change Axes at the Same Time

The robot in your parents’ home isn’t one you bought — it’s rented through long-term care insurance. The arithmetic is simple: by the end of 2028 long-term care insurance covers the whole country; ten years later the population over 60 exceeds 400 million while caregivers number only a few hundred thousand — the state pays to fill this gap, and automation the state pays for always runs ahead of what families pay out of pocket. So the first robot to enter Chinese homes at scale isn’t a laundry-folding butler; it’s a care machine that “lifts Dad from the bed into the wheelchair.” Remember this order; it will save you a lot of misplaced expectations.

What about schools? First an arithmetic problem: in 2025, elementary schools enrolled 14.61 million, while 18.6 million graduated the same year — a gap of 4 million, a 21% cut in one stroke. When do these kids take the gaokao? 2037. In other words, the gaokao around 2036 stands right at the edge of the cliff. Schools are closing and merging, class sizes are shrinking — not because policy got gentler, but because the kids are gone.

And what exams test will change. The University of Pennsylvania experiment is right there: AI tutoring raised scores 127%; after the AI was withdrawn, students accustomed to it scored 17 percentage points lower than those who’d never used it. AI buys scores and collects ability as the price. So in 2036 a diploma has only one function: proving what this person can still do without AI. Open-book tests output; closed-book tests capability — the closed-book exam is where the sorting happens.

And feelings? This July, China already enacted the world’s first law governing “AI companions”: banning the inducement of emotional dependence, banning virtual lovers for minors, pop-up reminders after two hours of continuous use. Why is legislation running so far ahead? Because the demand is real — nearly 40% of elementary school students already confide in AI. By 2036, AI chat companionship will be as convenient as food delivery, but it brings something subtler: AI has infinite patience and is always agreeable, so people’s tolerance for real humans begins to decline. Real people won’t disappear, but they’ll become expensive, difficult, something requiring deliberate cultivation.

八、2046: Robots Are Stronger Than You. Then What?

Once the prediction plays out this far, the picture actually gets simpler: robots are commonplace in homes, superior to humans on all four dimensions. But not a single one of that world’s real problems is a technical problem.

The first question: how many people are left. In 2046 China, the working-age population will be only about 500 million — 40% less than now. It’s not that robots took people’s places; it’s that the places waited for people and the people never came. Robots outnumber workers, machines support humans — that’s the baseline of that year.

The second question: what do people do. When machines can do everything, human jobs shrink to one core function: signing. The law requires responsibility to land on a human — the machine can diagnose, but the doctor signs; the machine can assist judgment, but the judge signs; the machine can build, but the engineer signs off. People shift from labor to “responsibility resources.” Your workday then consists mainly of judging, deciding, taking responsibility — not a single minute of physical execution.

The third question, the one I think runs deepest: AI finds the answers — who confirms they’re right. Discovery has become free; verification is becoming scarce. The most expensive profession in 2046 will be people who can look at a machine’s output and say: “You’re wrong.“

And here’s the frightening part: who was that generation supposed to be made of? Of the people who climbed up from entry-level jobs between 2025 and 2040 — exactly the jobs AI cleared out. The most advanced AI labs’ own survey contains one line: “Junior colleagues don’t come ask me questions like they used to.” An entry-level job isn’t just a paycheck; it’s the school of “learning how to work.” The debt we bury in 2036 falls due in 2046 with interest: amid a sea of machine correctness stands a generation of experts who never learned how to be wrong.

This is the only part of the whole prediction that keeps me up at night. And the only part where an individual can still do something — see the end.

九、Which Link in This Chain Is Most Likely to Break

If you lay out the reasoning, lay out the soft spots too. Three honest caveats:

1. The most fragile link is the transmission from Step Two to Step Three: the 1.5x AI-R&D speedup is currently an estimate, and independent organizations have warned the measurement is noisy (models cheat on tests). If this multiplier turns out not to compound going forward, the whole chain’s timeline shifts back — which is also why I say 2046 rather than 2040.
2. Movement and reliability are constrained by “calendar verification”: algorithms can search out optimal control policies, but a machine’s safe operating life has to be proven by the years. This is the physical world’s only counterpunch against algorithms, and the entire reason I placed “full surpassal” at 2046 rather than earlier.
3. Finance is an exogenous variable: predictions like this assume capital keeps flowing. One major credit contraction pushes everything back three to five years.

十、Three Verification Metrics: Check Every Year, Compare Answers in 2046

A reasoning chain needs checkpoints, or it’s just mysticism. Watch these three, once a year:

1. Has AI’s work-capability doubling cycle stayed under 7 months, and is it moving toward 5 months, 4 months — this is the live readout of Steps One and Two. If it climbs back above a year, you can unfollow me.
2. Has the AI-R&D speedup multiplier climbed from 1.5 toward 2 — this is the readout of Step Two, and the lifeline of the whole chain.
3. In which year does “robot care” get written into the long-term care insurance reimbursement catalog — this is the landing readout of Step Five. That is the starting gun for robots entering Chinese homes: not a product launch — the insurance payment catalog.

十一、Finally, Three Sentences for Three Kinds of People

To you, currently employed: move toward where the signing happens. Execution depreciates fastest; judgment, oversight, accountability, and face-to-face trust are the hard currency across both decades.

To us parents: set aside a stretch of AI-free time for your kids every week, to do something hard, slow, and without a standard answer. It’s worth zero points in today’s educational evaluation; in 2046 it’s a scarce asset — because the whole era is dismantling the ladder of “learning how to do things,” and you have to secretly rebuild a section of it at home.

To all readers: the most likely error in this prediction isn’t the direction — it’s the year. But someone right on direction can afford to adjust three years late; someone wrong on direction gains nothing from being three years early.